Synopsis: Discovering New Magnetic Materials with Machine Learning

A new computing experiment suggests that machine-learning algorithms can accelerate the discovery and design of new magnetic materials.

Data-storage technologies depend on materials that sustain magnetic properties at high temperature. While researchers have a range of such materials to work with, theory suggests that the known options are but a small fraction of the high-temperature magnets that are possible. To speed up the discovery and design of new high-temperature magnets, James Nelson and Stefano Sanvito of Trinity College in Ireland have developed several machine-learning models that can predict the temperature at which a material demagnetizes—its Curie temperature—from its chemical composition.

The researchers took empirical data from 2500 known ferromagnets and split them into two sets. The computer analyzed one set to build the predictive models and the other set to evaluate their accuracy. Each model describes the relationship between a material’s Curie temperature and several other properties, such as its atomic number, its melting temperature, and the type of bonds that form between the atoms. In most cases, they were able to predict a material’s Curie temperature from just its chemical formula.

Nelson and Sanvito found that the best model could predict a material’s Curie temperature with an accuracy of about 50 K. More importantly, this model could extrapolate from relatively few data points. For example, the model correctly predicted Curie-temperature changes for manganese-cobalt alloys as their composition varied, even though the training set contained only two data points for these materials. This approach to materials discovery still has important constraints, however. One limitation, for example, is that it cannot yet distinguish between polymorphs—compositionally identical materials whose Curie temperatures differ because of their distinct structures. This problem indicates that the model has some fundamental error that cannot be reduced by increasing the amount of training data, and means that, although the algorithm can accelerate the materials design process, researchers need other methods to confirm the predicted properties.

This research is published in Physical Review Materials.

–Sophia Chen

Sophia Chen is a freelance science writer based in Tucson, Arizona.


Features

More Features »

Announcements

More Announcements »

Subject Areas

Materials ScienceMagnetism

Previous Synopsis

Next Synopsis

Related Articles

Synopsis: Soft Biological Tissues Can Be Piezoelectric
Biological Physics

Synopsis: Soft Biological Tissues Can Be Piezoelectric

Artery walls, tendons, and heart valves can generate an electric voltage when squeezed—an effect that could be harnessed to diagnose important diseases. Read More »

Viewpoint: Phonon Heat Transport Near the Melting Point
Materials Science

Viewpoint: Phonon Heat Transport Near the Melting Point

Molecular dynamics simulations can fully describe phonon propagation in aluminum, which could enable accurate predictions of phonon thermal conductivity. Read More »

Viewpoint: The Heat in Antiferromagnetic Switching
Condensed Matter Physics

Viewpoint: The Heat in Antiferromagnetic Switching

New experiments suggest that heat might be responsible for the current-induced voltage signals measured in antiferromagnets, and not a rotation of the material’s spins as previously thought. Read More »

More Articles